use std::arch::x86_64::{_mm_loadu_si128, __m128i, _mm_cmpeq_epi8, _mm_set1_epi8, _mm_sub_epi8, _mm_and_si128, _mm_unpackhi_epi8, _mm_add_epi32, _mm_shufflelo_epi16};
//use std::arch::x86::_MM_SHUFFLE;
use std::arch::x86;

use std::mem;

fn main() {
    unsafe {
        //let t = _mm_set1_epi8(50);
        let q = _mm_set1_epi8(51);


        let hi8 = _mm_unpackhi_epi8(q, q);

        let sum64 = _mm_add_epi32(hi8, q);

        let hi32  = _mm_shufflelo_epi16(sum64, _MM_SHUFFLE(1, 0, 3, 2));

        let arr: [u8; 16] = mem::transmute(hi32);

        println!("{:?}", arr);
    }

    println!("{} {}", 255 & 48, 255 - 48);
}

/*
    255 - 50 = 205
    50 == 50 -> 255

    207 - 205 = 2





*/